CCPNet: Joining the pooling transformer and target context for medical image segmentation

文献类型： 外文期刊

第一作者： Yang, Yakun

作者： Yang, Yakun;Xue, Hongcheng;Wang, Longhe;Li, Lin;Liu, Xiangping;Feng, Chungang;Liu, Xiangping;Qu, Hao;Qu, Hao

作者机构：

关键词： Computer vision; Medical image segmentation; Pooling transformer; Target context; Tibial dyschondroplasia analysis

期刊名称：PATTERN RECOGNITION （ 影响因子：7.6； 五年影响因子：7.9 ）

ISSN： 0031-3203

年卷期： 2026 年 171 卷

页码：

收录情况： SCI

摘要： The automated and precise segmentation of medical images is vital for the clinical diagnosis and treatment planning. Recent segmentation networks utilize the self-attention that calculates a global similarity matrix to capture the long-range dependency, which breaks the local limitation in the convolution. However, the quadratic complexity is unbearable when integrating the global context of high-resolution features. Besides, existing researches prefer to propose powerful encoders but neglect to design ingenious decoders. Following the encoder structure to design decoder is difficult to generate precise masks, owing to both difference in functionality. In this paper, we propose a segmentation network CCPNet to avoid mentioned drawbacks. Specifically, we redesign the encoding unit to efficiently blend details and semantics in each scale, and optimize its structure from the standpoint of maximizing the gradient combination. Moreover, we design the decoder that utilizes prior intermediate probability to formulate target contexts and refine decoding features through the spatial reduction cross-attention. Overall, it forms cascade adjustments to enhance the inter-category consistency and the intra-category difference. Our extensive evaluations on three public benchmarks and one self-built dataset, reveal the superiority of our CCPNet in terms of accuracy and complexity. On ACT-1K, our approach obtains best accuracy(76.38 % on mIoU and 85.04% on mDice) compared to existing typical methods. In applications, we originally establish a standard to quantitatively analyze the tibial dyschondroplasia based on the segmented masks of self-built dataset. The code will be available at https://github.com/LATIESUS/CCPNet.

分类号：

  • 相关文献
作者其他论文 更多>>

微信小程序